[DAGCombiner][X86] Improve narrowExtractedVectorLoad to handle cases where the element size isn't byte sized by the subvector is.

[craig.topper]

Summary

Follow up from D75377. If the subvector is byte sized and the
index is aligned to the subvector size, we can shrink the load.

[RKSimon]

LGTM - thanks Craig

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

static SDValue narrowExtractedVectorLoad(SDNode *Extract, SelectionDAG &DAG) {
   auto *ExtIdx = dyn_cast<ConstantSDNode>(Extract->getOperand(1));
   if (!Ld || Ld->getExtensionType() || !Ld->isSimple() ||
       !ExtIdx)
     return SDValue();
 
   // Allow targets to opt-out.
   EVT VT = Extract->getValueType(0);
 
-  // Only handle byte sized scalars otherwise the offset is incorrect.
-  // FIXME: We might be able to do better if the VT is byte sized and the index
-  // is aligned.
-  if (!VT.getScalarType().isByteSized())
+  // We can only create byte sized loads.
+  if (!VT.isByteSized())
+    return SDValue();
+
+  unsigned Index = ExtIdx->getZExtValue();
+  unsigned NumElts = VT.getVectorNumElements();
+
+  // If the index is a multiple of the extract element count, we can offset the
+  // address by the store size multiplied by the subvector index. Otherwise if
+  // the scalar type is byte sized, we can just use the index multiplied by
+  // the element size in bytes as the offset.
+  unsigned Offset;
+  if (Index % NumElts == 0)
+    Offset = (Index / NumElts) * VT.getStoreSize();
+  else if (VT.getScalarType().isByteSized())
+    Offset = Index * VT.getScalarType().getStoreSize();
+  else
     return SDValue();
 
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   if (!TLI.shouldReduceLoadWidth(Ld, Ld->getExtensionType(), VT))
     return SDValue();
 
   // The narrow load will be offset from the base address of the old load if
   // we are extracting from something besides index 0 (little-endian).
   SDLoc DL(Extract);
-  SDValue BaseAddr = Ld->getOperand(1);
-  unsigned Offset = ExtIdx->getZExtValue() * VT.getScalarType().getStoreSize();
+  SDValue BaseAddr = Ld->getBasePtr();
 
   // TODO: Use "BaseIndexOffset" to make this more effective.
   SDValue NewAddr = DAG.getMemBasePlusOffset(BaseAddr, Offset, DL);
   MachineFunction &MF = DAG.getMachineFunction();
   MachineMemOperand *MMO = MF.getMachineMemOperand(Ld->getMemOperand(), Offset,
                                                    VT.getStoreSize());
   SDValue NewLd = DAG.getLoad(VT, DL, Ld->getChain(), NewAddr, MMO);
   DAG.makeEquivalentMemoryOrdering(Ld, NewLd);

llvm/test/CodeGen/X86/avx512-extract-subvector-load-store.ll

; AVX512NOTDQ-NEXT:    retq
     %d1 = shufflevector <32 x i1> %d0,<32 x i1> undef,<4 x i32><i32 16,i32 16,i32 16,i32 16>
     %d2 = select <4 x i1> %d1, <4 x float> %a1, <4 x float> %a2
     store <4 x float> %d2, <4 x float>* %a3
     ret void
 }
 define void @load_v32i1_broadcast_16_v8i1(<32 x i1>* %a0,<8 x float> %a1,<8 x float> %a2,<8 x float>* %a3) {
 ; AVX512-LABEL: load_v32i1_broadcast_16_v8i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovd (%rdi), %k0
-; AVX512-NEXT:    kshiftrd $16, %k0, %k0
+; AVX512-NEXT:    kmovb 2(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm2
 ; AVX512-NEXT:    vpbroadcastd %xmm2, %ymm2
 ; AVX512-NEXT:    vpmovd2m %ymm2, %k1
 ; AVX512-NEXT:    vmovaps %ymm0, %ymm1 {%k1}
 ; AVX512-NEXT:    vmovaps %ymm1, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_16_v8i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrd $16, %k0, %k1
+; AVX512NOTDQ-NEXT:    kmovw 2(%rdi), %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm2, %ymm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vptestmd %ymm2, %ymm2, %k1
 ; AVX512NOTDQ-NEXT:    vmovaps %ymm0, %ymm1 {%k1}
 ; AVX512NOTDQ-NEXT:    vmovaps %ymm1, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
; AVX512NOTDQ-NEXT:    retq
     %d1 = shufflevector <32 x i1> %d0,<32 x i1> undef,<4 x i32><i32 31,i32 31,i32 31,i32 31>
     %d2 = select <4 x i1> %d1, <4 x float> %a1, <4 x float> %a2
     store <4 x float> %d2, <4 x float>* %a3
     ret void
 }
 define void @load_v32i1_broadcast_31_v8i1(<32 x i1>* %a0,<8 x float> %a1,<8 x float> %a2,<8 x float>* %a3) {
 ; AVX512-LABEL: load_v32i1_broadcast_31_v8i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovd (%rdi), %k0
-; AVX512-NEXT:    kshiftrd $24, %k0, %k0
+; AVX512-NEXT:    kmovb 3(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm2
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [7,7,7,7,7,7,7,7]
 ; AVX512-NEXT:    vpermd %ymm2, %ymm3, %ymm2
 ; AVX512-NEXT:    vpmovd2m %ymm2, %k1
 ; AVX512-NEXT:    vmovaps %ymm0, %ymm1 {%k1}
 ; AVX512-NEXT:    vmovaps %ymm1, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_31_v8i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrd $24, %k0, %k1
+; AVX512NOTDQ-NEXT:    movzbl 3(%rdi), %eax
+; AVX512NOTDQ-NEXT:    kmovd %eax, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm2, %ymm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [7,7,7,7,7,7,7,7]
 ; AVX512NOTDQ-NEXT:    vpermd %ymm2, %ymm3, %ymm2
 ; AVX512NOTDQ-NEXT:    vptestmd %ymm2, %ymm2, %k1
 ; AVX512NOTDQ-NEXT:    vmovaps %ymm0, %ymm1 {%k1}
 ; AVX512NOTDQ-NEXT:    vmovaps %ymm1, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
; AVX512NOTDQ-NEXT:    retq
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<4 x i32><i32 32,i32 32,i32 32,i32 32>
     %d2 = select <4 x i1> %d1, <4 x float> %a1, <4 x float> %a2
     store <4 x float> %d2, <4 x float>* %a3
     ret void
 }
 define void @load_v64i1_broadcast_32_v8i1(<64 x i1>* %a0,<8 x float> %a1,<8 x float> %a2,<8 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v8i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovq (%rdi), %k0
-; AVX512-NEXT:    kshiftrq $32, %k0, %k0
+; AVX512-NEXT:    kmovb 4(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm2
 ; AVX512-NEXT:    vpbroadcastd %xmm2, %ymm2
 ; AVX512-NEXT:    vpmovd2m %ymm2, %k1
 ; AVX512-NEXT:    vmovaps %ymm0, %ymm1 {%k1}
 ; AVX512-NEXT:    vmovaps %ymm1, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v8i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
+; AVX512NOTDQ-NEXT:    kmovw 4(%rdi), %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm2, %ymm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vptestmd %ymm2, %ymm2, %k1
 ; AVX512NOTDQ-NEXT:    vmovaps %ymm0, %ymm1 {%k1}
 ; AVX512NOTDQ-NEXT:    vmovaps %ymm1, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<8 x i32><i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32>
     %d2 = select <8 x i1> %d1, <8 x float> %a1, <8 x float> %a2
     store <8 x float> %d2, <8 x float>* %a3
     ret void
 }
 define void @load_v64i1_broadcast_32_v16i1(<64 x i1>* %a0,<16 x float> %a1,<16 x float> %a2,<16 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v16i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovq (%rdi), %k0
-; AVX512-NEXT:    kshiftrq $32, %k0, %k0
+; AVX512-NEXT:    kmovw 4(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %zmm2
 ; AVX512-NEXT:    vpbroadcastd %xmm2, %zmm2
 ; AVX512-NEXT:    vpmovd2m %zmm2, %k1
 ; AVX512-NEXT:    vmovaps %zmm0, %zmm1 {%k1}
 ; AVX512-NEXT:    vmovaps %zmm1, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v16i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
+; AVX512NOTDQ-NEXT:    kmovw 4(%rdi), %k1
 ; AVX512NOTDQ-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm2, %zmm2
 ; AVX512NOTDQ-NEXT:    vptestmd %zmm2, %zmm2, %k1
 ; AVX512NOTDQ-NEXT:    vmovaps %zmm0, %zmm1 {%k1}
 ; AVX512NOTDQ-NEXT:    vmovaps %zmm1, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
; AVX512NOTDQ-NEXT:    retq
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<4 x i32><i32 63,i32 63,i32 63,i32 63>
     %d2 = select <4 x i1> %d1, <4 x float> %a1, <4 x float> %a2
     store <4 x float> %d2, <4 x float>* %a3
     ret void
 }
 define void @load_v64i1_broadcast_63_v8i1(<64 x i1>* %a0,<8 x float> %a1,<8 x float> %a2,<8 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v8i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovq (%rdi), %k0
-; AVX512-NEXT:    kshiftrq $56, %k0, %k0
+; AVX512-NEXT:    kmovb 7(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm2
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [7,7,7,7,7,7,7,7]
 ; AVX512-NEXT:    vpermd %ymm2, %ymm3, %ymm2
 ; AVX512-NEXT:    vpmovd2m %ymm2, %k1
 ; AVX512-NEXT:    vmovaps %ymm0, %ymm1 {%k1}
 ; AVX512-NEXT:    vmovaps %ymm1, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v8i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrq $56, %k0, %k1
+; AVX512NOTDQ-NEXT:    movzbl 7(%rdi), %eax
+; AVX512NOTDQ-NEXT:    kmovd %eax, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm2, %ymm2, %ymm2
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm2, %ymm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} ymm3 = [7,7,7,7,7,7,7,7]
 ; AVX512NOTDQ-NEXT:    vpermd %ymm2, %ymm3, %ymm2
 ; AVX512NOTDQ-NEXT:    vptestmd %ymm2, %ymm2, %k1
 ; AVX512NOTDQ-NEXT:    vmovaps %ymm0, %ymm1 {%k1}
 ; AVX512NOTDQ-NEXT:    vmovaps %ymm1, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<8 x i32><i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63>
     %d2 = select <8 x i1> %d1, <8 x float> %a1, <8 x float> %a2
     store <8 x float> %d2, <8 x float>* %a3
     ret void
 }
 define void @load_v64i1_broadcast_63_v16i1(<64 x i1>* %a0,<16 x float> %a1,<16 x float> %a2,<16 x float>* %a3) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v16i1:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovq (%rdi), %k0
-; AVX512-NEXT:    kshiftrq $48, %k0, %k0
+; AVX512-NEXT:    kmovw 6(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %zmm2
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX512-NEXT:    vpermd %zmm2, %zmm3, %zmm2
 ; AVX512-NEXT:    vpmovd2m %zmm2, %k1
 ; AVX512-NEXT:    vmovaps %zmm0, %zmm1 {%k1}
 ; AVX512-NEXT:    vmovaps %zmm1, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v16i1:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrq $48, %k0, %k1
+; AVX512NOTDQ-NEXT:    kmovw 6(%rdi), %k1
 ; AVX512NOTDQ-NEXT:    vpternlogd $255, %zmm2, %zmm2, %zmm2 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} zmm3 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX512NOTDQ-NEXT:    vpermd %zmm2, %zmm3, %zmm2
 ; AVX512NOTDQ-NEXT:    vptestmd %zmm2, %zmm2, %k1
 ; AVX512NOTDQ-NEXT:    vmovaps %zmm0, %zmm1 {%k1}
 ; AVX512NOTDQ-NEXT:    vmovaps %zmm1, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
; AVX512NOTDQ-NEXT:    retq
     %d0 = load <32 x i1>, <32 x i1>* %a0
     %d1 = shufflevector <32 x i1> %d0,<32 x i1> undef,<4 x i32><i32 16,i32 16,i32 16,i32 16>
     store <4 x i1> %d1, <4 x i1>* %a1
     ret void
 }
 define void @load_v32i1_broadcast_16_v8i1_store(<32 x i1>* %a0,<8 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_16_v8i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovd (%rdi), %k0
-; AVX512-NEXT:    kshiftrd $16, %k0, %k0
+; AVX512-NEXT:    kmovb 2(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm0
 ; AVX512-NEXT:    vpbroadcastd %xmm0, %ymm0
 ; AVX512-NEXT:    vpmovd2m %ymm0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_16_v8i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrd $16, %k0, %k1
+; AVX512NOTDQ-NEXT:    kmovw 2(%rdi), %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vptestmd %ymm0, %ymm0, %k0
 ; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
; AVX512NOTDQ-NEXT:    retq
     %d0 = load <32 x i1>, <32 x i1>* %a0
     %d1 = shufflevector <32 x i1> %d0,<32 x i1> undef,<4 x i32><i32 31,i32 31,i32 31,i32 31>
     store <4 x i1> %d1, <4 x i1>* %a1
     ret void
 }
 define void @load_v32i1_broadcast_31_v8i1_store(<32 x i1>* %a0,<8 x i1>* %a1) {
 ; AVX512-LABEL: load_v32i1_broadcast_31_v8i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovd (%rdi), %k0
-; AVX512-NEXT:    kshiftrd $24, %k0, %k0
+; AVX512-NEXT:    kmovb 3(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm0
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [7,7,7,7,7,7,7,7]
 ; AVX512-NEXT:    vpermd %ymm0, %ymm1, %ymm0
 ; AVX512-NEXT:    vpmovd2m %ymm0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v32i1_broadcast_31_v8i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovd (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrd $24, %k0, %k1
+; AVX512NOTDQ-NEXT:    movzbl 3(%rdi), %eax
+; AVX512NOTDQ-NEXT:    kmovd %eax, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [7,7,7,7,7,7,7,7]
 ; AVX512NOTDQ-NEXT:    vpermd %ymm0, %ymm1, %ymm0
 ; AVX512NOTDQ-NEXT:    vptestmd %ymm0, %ymm0, %k0
 ; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<4 x i32><i32 32,i32 32,i32 32,i32 32>
     store <4 x i1> %d1, <4 x i1>* %a1
     ret void
 }
 define void @load_v64i1_broadcast_32_v8i1_store(<64 x i1>* %a0,<8 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v8i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovq (%rdi), %k0
-; AVX512-NEXT:    kshiftrq $32, %k0, %k0
+; AVX512-NEXT:    kmovb 4(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm0
 ; AVX512-NEXT:    vpbroadcastd %xmm0, %ymm0
 ; AVX512-NEXT:    vpmovd2m %ymm0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v8i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
+; AVX512NOTDQ-NEXT:    kmovw 4(%rdi), %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vptestmd %ymm0, %ymm0, %k0
 ; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<8 x i32><i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32>
     store <8 x i1> %d1, <8 x i1>* %a1
     ret void
 }
 define void @load_v64i1_broadcast_32_v16i1_store(<64 x i1>* %a0,<16 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_32_v16i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovq (%rdi), %k0
-; AVX512-NEXT:    kshiftrq $32, %k0, %k0
+; AVX512-NEXT:    kmovw 4(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %zmm0
 ; AVX512-NEXT:    vpbroadcastd %xmm0, %zmm0
 ; AVX512-NEXT:    vpmovd2m %zmm0, %k0
 ; AVX512-NEXT:    kmovw %k0, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_32_v16i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrq $32, %k0, %k1
+; AVX512NOTDQ-NEXT:    kmovw 4(%rdi), %k1
 ; AVX512NOTDQ-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd %xmm0, %zmm0
 ; AVX512NOTDQ-NEXT:    vptestmd %zmm0, %zmm0, %k0
 ; AVX512NOTDQ-NEXT:    kmovw %k0, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<16 x i32><i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32,i32 32>
; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<4 x i32><i32 63,i32 63,i32 63,i32 63>
     store <4 x i1> %d1, <4 x i1>* %a1
     ret void
 }
 define void @load_v64i1_broadcast_63_v8i1_store(<64 x i1>* %a0,<8 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v8i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovq (%rdi), %k0
-; AVX512-NEXT:    kshiftrq $56, %k0, %k0
+; AVX512-NEXT:    kmovb 7(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %ymm0
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [7,7,7,7,7,7,7,7]
 ; AVX512-NEXT:    vpermd %ymm0, %ymm1, %ymm0
 ; AVX512-NEXT:    vpmovd2m %ymm0, %k0
 ; AVX512-NEXT:    kmovb %k0, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v8i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrq $56, %k0, %k1
+; AVX512NOTDQ-NEXT:    movzbl 7(%rdi), %eax
+; AVX512NOTDQ-NEXT:    kmovd %eax, %k1
 ; AVX512NOTDQ-NEXT:    vpcmpeqd %ymm0, %ymm0, %ymm0
 ; AVX512NOTDQ-NEXT:    vmovdqa32 %ymm0, %ymm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [7,7,7,7,7,7,7,7]
 ; AVX512NOTDQ-NEXT:    vpermd %ymm0, %ymm1, %ymm0
 ; AVX512NOTDQ-NEXT:    vptestmd %ymm0, %ymm0, %k0
 ; AVX512NOTDQ-NEXT:    kmovd %k0, %eax
 ; AVX512NOTDQ-NEXT:    movb %al, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<8 x i32><i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63>
     store <8 x i1> %d1, <8 x i1>* %a1
     ret void
 }
 define void @load_v64i1_broadcast_63_v16i1_store(<64 x i1>* %a0,<16 x i1>* %a1) {
 ; AVX512-LABEL: load_v64i1_broadcast_63_v16i1_store:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    kmovq (%rdi), %k0
-; AVX512-NEXT:    kshiftrq $48, %k0, %k0
+; AVX512-NEXT:    kmovw 6(%rdi), %k0
 ; AVX512-NEXT:    vpmovm2d %k0, %zmm0
 ; AVX512-NEXT:    vpbroadcastd {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX512-NEXT:    vpermd %zmm0, %zmm1, %zmm0
 ; AVX512-NEXT:    vpmovd2m %zmm0, %k0
 ; AVX512-NEXT:    kmovw %k0, (%rsi)
 ; AVX512-NEXT:    vzeroupper
 ; AVX512-NEXT:    retq
 ;
 ; AVX512NOTDQ-LABEL: load_v64i1_broadcast_63_v16i1_store:
 ; AVX512NOTDQ:       # %bb.0:
-; AVX512NOTDQ-NEXT:    kmovq (%rdi), %k0
-; AVX512NOTDQ-NEXT:    kshiftrq $48, %k0, %k1
+; AVX512NOTDQ-NEXT:    kmovw 6(%rdi), %k1
 ; AVX512NOTDQ-NEXT:    vpternlogd $255, %zmm0, %zmm0, %zmm0 {%k1} {z}
 ; AVX512NOTDQ-NEXT:    vpbroadcastd {{.*#+}} zmm1 = [15,15,15,15,15,15,15,15,15,15,15,15,15,15,15,15]
 ; AVX512NOTDQ-NEXT:    vpermd %zmm0, %zmm1, %zmm0
 ; AVX512NOTDQ-NEXT:    vptestmd %zmm0, %zmm0, %k0
 ; AVX512NOTDQ-NEXT:    kmovw %k0, (%rsi)
 ; AVX512NOTDQ-NEXT:    vzeroupper
 ; AVX512NOTDQ-NEXT:    retq
     %d0 = load <64 x i1>, <64 x i1>* %a0
     %d1 = shufflevector <64 x i1> %d0,<64 x i1> undef,<16 x i32><i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63,i32 63>
     store <16 x i1> %d1, <16 x i1>* %a1
     ret void
 }
 

llvm/test/CodeGen/X86/avx512-mask-op.ll

 ; AVX512DQ-NEXT:    kmovw %edi, %k0
 ; AVX512DQ-NEXT:    kshiftrw $8, %k0, %k0
 ; AVX512DQ-NEXT:    kmovw %k0, %eax
 ; AVX512DQ-NEXT:    ## kill: def $al killed $al killed $eax
 ; AVX512DQ-NEXT:    retq
 ;
 ; X86-LABEL: shuf_test1:
 ; X86:       ## %bb.0:
-; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k0
-; X86-NEXT:    kshiftrw $8, %k0, %k0
-; X86-NEXT:    kmovd %k0, %eax
-; X86-NEXT:    ## kill: def $al killed $al killed $eax
+; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
 ; X86-NEXT:    retl
    %v1 = bitcast i16 %v to <16 x i1>
    %mask = shufflevector <16 x i1> %v1, <16 x i1> undef, <8 x i32> <i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15>
    %mask1 = bitcast <8 x i1> %mask to i8
    ret i8 %mask1
 }
 
 define i32 @zext_test1(<16 x i32> %a, <16 x i32> %b) {

llvm/test/CodeGen/X86/avx512bw-intrinsics-upgrade.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512bw --show-mc-encoding | FileCheck %s --check-prefixes=CHECK,X86
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512bw --show-mc-encoding | FileCheck %s --check-prefixes=CHECK,X64
 
 declare i32 @llvm.x86.avx512.kunpck.wd(i32, i32)
 
 define i32@test_int_x86_avx512_kunpck_wd(i32 %x0, i32 %x1) {
 ; X86-LABEL: test_int_x86_avx512_kunpck_wd:
 ; X86:       # %bb.0:
-; X86-NEXT:    kmovd {{[0-9]+}}(%esp), %k0 # encoding: [0xc4,0xe1,0xf9,0x90,0x44,0x24,0x04]
-; X86-NEXT:    kmovd {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf9,0x90,0x4c,0x24,0x08]
+; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k0 # encoding: [0xc5,0xf8,0x90,0x44,0x24,0x04]
+; X86-NEXT:    kmovw {{[0-9]+}}(%esp), %k1 # encoding: [0xc5,0xf8,0x90,0x4c,0x24,0x08]
 ; X86-NEXT:    kunpckwd %k1, %k0, %k0 # encoding: [0xc5,0xfc,0x4b,0xc1]
 ; X86-NEXT:    kmovd %k0, %eax # encoding: [0xc5,0xfb,0x93,0xc0]
 ; X86-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-LABEL: test_int_x86_avx512_kunpck_wd:
 ; X64:       # %bb.0:
 ; X64-NEXT:    kmovd %edi, %k0 # encoding: [0xc5,0xfb,0x92,0xc7]
 ; X64-NEXT:    kmovd %esi, %k1 # encoding: [0xc5,0xfb,0x92,0xce]
 ; X64-NEXT:    kunpckwd %k1, %k0, %k0 # encoding: [0xc5,0xfc,0x4b,0xc1]
 ; X64-NEXT:    kmovd %k0, %eax # encoding: [0xc5,0xfb,0x93,0xc0]
 ; X64-NEXT:    retq # encoding: [0xc3]
   %res = call i32 @llvm.x86.avx512.kunpck.wd(i32 %x0, i32 %x1)
   ret i32 %res
 }
 
 declare i64 @llvm.x86.avx512.kunpck.dq(i64, i64)
 
 define i64@test_int_x86_avx512_kunpck_qd(i64 %x0, i64 %x1) {
 ; X86-LABEL: test_int_x86_avx512_kunpck_qd:
 ; X86:       # %bb.0:
-; X86-NEXT:    kmovq {{[0-9]+}}(%esp), %k0 # encoding: [0xc4,0xe1,0xf8,0x90,0x44,0x24,0x04]
-; X86-NEXT:    kmovq {{[0-9]+}}(%esp), %k1 # encoding: [0xc4,0xe1,0xf8,0x90,0x4c,0x24,0x0c]
-; X86-NEXT:    kmovd %k0, %edx # encoding: [0xc5,0xfb,0x93,0xd0]
-; X86-NEXT:    kmovd %k1, %eax # encoding: [0xc5,0xfb,0x93,0xc1]
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx # encoding: [0x8b,0x54,0x24,0x04]
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax # encoding: [0x8b,0x44,0x24,0x0c]
 ; X86-NEXT:    retl # encoding: [0xc3]
 ;
 ; X64-LABEL: test_int_x86_avx512_kunpck_qd:
 ; X64:       # %bb.0:
 ; X64-NEXT:    kmovq %rdi, %k0 # encoding: [0xc4,0xe1,0xfb,0x92,0xc7]
 ; X64-NEXT:    kmovq %rsi, %k1 # encoding: [0xc4,0xe1,0xfb,0x92,0xce]
 ; X64-NEXT:    kunpckdq %k1, %k0, %k0 # encoding: [0xc4,0xe1,0xfc,0x4b,0xc1]
 ; X64-NEXT:    kmovq %k0, %rax # encoding: [0xc4,0xe1,0xfb,0x93,0xc0]